aprict 0.0.23__tar.gz → 0.0.24.dev1__tar.gz

{aprict-0.0.23/src/aprict.egg-info → aprict-0.0.24.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aprict
-Version: 0.0.23
+Version: 0.0.24.dev1
 Summary: Tools to simulate and reconstruct XCT transmission data.
 Author: Murdock Grewar
 Author-email: Murdock Grewar <murdock.grewar@anu.edu.au>
@@ -36,6 +36,10 @@ Requires-Dist: neuroglancer>=2.41.0; extra == "gui"
 Dynamic: author
 Dynamic: license-file
 
+[![Static Badge](https://img.shields.io/badge/ORCiD-0000--0002--0015--8284-green?logo=orcid)](https://orcid.org/0000-0002-0015-8284)
+[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.23632625.svg)](https://doi.org/10.5281/zenodo.23632625)
+![PyPI Version](https://img.shields.io/pypi/v/aprict)
+[![Project Status: WIP – Initial development is in progress, but there has not yet been a stable, usable release suitable for the public.](https://www.repostatus.org/badges/latest/wip.svg)](https://www.repostatus.org/#wip)
 <details open><summary><h1>About</h1></summary>
 
 Description
@@ -71,8 +75,7 @@ Known limitations
   * Volume size is limited by CUDA texture limit. For example, a `(2400,2400,4800)` volume can be reconstructed, but not a `(30000,30000,1)` volume, due to a volume of this size not fitting within a single CUDA texture
   * The FORBILD shape `Tetrahedron` is not yet implemented, and no primitive shape is permitted to have more than 3 clipping planes applied, due to internal limitations. All other FORBILD phantoms should be supported.
   * Only one GPU is utilised by this module, at the present time. There is no defined behaviour for which GPU will be chosen.
-  * Not fully optimised: among other issues, 
-    * the GBC reconstruction script assumes that the whole volume (or more) can fit in RAM. This limitation may be lifted in the future. (Note: a present workaround is to create a large swapfile on disk, since swap space is treated as RAM when the physical RAM is exhausted.)
+  * Not fully optimised
   * CPU and GPU work not completely asynchronous: due to unresolved bugs, various asynchronous GPU operations have been disabled. This may not have much impact on large reconstructions, but could result in poor GPU utilisation for smaller reconstructions.
   * No smart batching of (back)projections: smarter code would analyse the geometry of projections, and batch together (back)projections that involve similar regions of the volume. Currently, the code does not do any kind of clever batching like this, and so the (back)projection code is not as performant as it could be.
 
@@ -99,7 +102,7 @@ You can install `aprict` like you would any python package, using `pip install`.
 Example installation using `uv`: 
 ```bash
 $ uv venv -p 3.14
-$ uv pip install aprict[gui]==0.0.19
+$ uv pip install aprict[gui]==0.0.24
 $ source .venv/bin/activate
 ```
 You should now have several executables available on `$PATH`. Try `aprict-cli --help` to see the polished commandline interface. 
@@ -118,11 +121,11 @@ Activate the environment:
 ```bash
 source .venv/bin/activate
 ```
-Now, try the shell command `aprict-cli --help`. Here is the output as at `v0.0.19`:
+Now, try the shell command `aprict-cli --help`. Here is the output as at `v0.0.24`:
 ```
 Usage: aprict-cli [OPTIONS] COMMAND [ARGS]...
 
-  Entry point for aprict CLI.
+  Entry point for CLI interface to aprict version 0.0.24
 
 Options:
   --version   Show the version and exit.
@@ -170,8 +173,8 @@ aprict-cli project cbfp \
     --output projs \
     --plot-mode PYQTGRAPH
 ````
-(Time on my machine: ~2 minutes, 55 seconds) 
-The option `--plot-mode PYQTGRAPH` will display a plot window of the projection data after it is created. (This will not be available if the module was not installed with the `[gui]` option.)
+(Time on my machine: ~2 minutes, 56 seconds) 
+The options `--ssu 3 --ssv 3` indicate that the line attenuation for each detector pixel is to be calculated as the average of a 3-by-3 grid of exact line integrals. This kind of improved accuracy is necessary for some phantoms, including this one, if using a large detector pixel size. The option `--plot-mode PYQTGRAPH` will display a plot window of the projection data after it is created. (This will not be available if the module was not installed with the `[gui]` option.)
 
 Now, let's reconstruct with three different algorithms:
  1. Global Backprojection-Convolution (GBC)
@@ -186,10 +189,14 @@ aprict-cli recon gbc-cone-cyl \
     --voxwid 0.5 \
     --volshape '(525,525,525)' \
     --alg GREWAR \
+    --f1 1.2 \
+    --f2 6 \
+    --n 9 \
+    --eps 1e-9 \
     --output recon_gbc \
     --plot-mode PYQTGRAPH
 ```
-(Time on my machine: ~37 seconds)
+(Time on my machine: ~39 seconds)
 
 To perform an FBP reconstruction, we do:
 ```bash
@@ -202,7 +209,7 @@ aprict-cli recon fbp-cone-cyl \
     --output recon_fbp \
     --plot-mode PYQTGRAPH
 ```
-(Time on my machine: ~12 seconds)
+(Time on my machine: ~19 seconds)
 
 To perform an FBP-VM reconstruction, we change `--alg` to `NOOGREWAR`:
 ```bash
@@ -215,7 +222,7 @@ aprict-cli recon fbp-cone-cyl \
     --output recon_fbp-vw \
     --plot-mode PYQTGRAPH
 ```
-(Time on my machine: ~22 seconds)
+(Time on my machine: ~26 seconds)
 
 
 </details>
@@ -253,7 +260,7 @@ aprict-cli project cbfp \
     --output projs \
     --plot-mode PYQTGRAPH
 ````
-(Time on my machine: ~1 minute, 46 seconds) 
+(Time on my machine: ~1 minute, 48 seconds) 
 The option `--plot-mode PYQTGRAPH` will display a plot window of the projection data (a sinogram) after it is created. (This will not be available if the module was not installed with the `[gui]` option.) 
 
 You will find a copy of the `ROCK_CAN` phantom in the output directory `projs`, beside the projection data. The phantom is always copied to the output directory, to ensure a good record of the inputs to the simulation.

{aprict-0.0.23 → aprict-0.0.24.dev1}/README.md

@@ -1,3 +1,7 @@
+[![Static Badge](https://img.shields.io/badge/ORCiD-0000--0002--0015--8284-green?logo=orcid)](https://orcid.org/0000-0002-0015-8284)
+[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.23632625.svg)](https://doi.org/10.5281/zenodo.23632625)
+![PyPI Version](https://img.shields.io/pypi/v/aprict)
+[![Project Status: WIP – Initial development is in progress, but there has not yet been a stable, usable release suitable for the public.](https://www.repostatus.org/badges/latest/wip.svg)](https://www.repostatus.org/#wip)
 <details open><summary><h1>About</h1></summary>
 
 Description
@@ -33,8 +37,7 @@ Known limitations
   * Volume size is limited by CUDA texture limit. For example, a `(2400,2400,4800)` volume can be reconstructed, but not a `(30000,30000,1)` volume, due to a volume of this size not fitting within a single CUDA texture
   * The FORBILD shape `Tetrahedron` is not yet implemented, and no primitive shape is permitted to have more than 3 clipping planes applied, due to internal limitations. All other FORBILD phantoms should be supported.
   * Only one GPU is utilised by this module, at the present time. There is no defined behaviour for which GPU will be chosen.
-  * Not fully optimised: among other issues, 
-    * the GBC reconstruction script assumes that the whole volume (or more) can fit in RAM. This limitation may be lifted in the future. (Note: a present workaround is to create a large swapfile on disk, since swap space is treated as RAM when the physical RAM is exhausted.)
+  * Not fully optimised
   * CPU and GPU work not completely asynchronous: due to unresolved bugs, various asynchronous GPU operations have been disabled. This may not have much impact on large reconstructions, but could result in poor GPU utilisation for smaller reconstructions.
   * No smart batching of (back)projections: smarter code would analyse the geometry of projections, and batch together (back)projections that involve similar regions of the volume. Currently, the code does not do any kind of clever batching like this, and so the (back)projection code is not as performant as it could be.
 
@@ -61,7 +64,7 @@ You can install `aprict` like you would any python package, using `pip install`.
 Example installation using `uv`: 
 ```bash
 $ uv venv -p 3.14
-$ uv pip install aprict[gui]==0.0.19
+$ uv pip install aprict[gui]==0.0.24
 $ source .venv/bin/activate
 ```
 You should now have several executables available on `$PATH`. Try `aprict-cli --help` to see the polished commandline interface. 
@@ -80,11 +83,11 @@ Activate the environment:
 ```bash
 source .venv/bin/activate
 ```
-Now, try the shell command `aprict-cli --help`. Here is the output as at `v0.0.19`:
+Now, try the shell command `aprict-cli --help`. Here is the output as at `v0.0.24`:
 ```
 Usage: aprict-cli [OPTIONS] COMMAND [ARGS]...
 
-  Entry point for aprict CLI.
+  Entry point for CLI interface to aprict version 0.0.24
 
 Options:
   --version   Show the version and exit.
@@ -132,8 +135,8 @@ aprict-cli project cbfp \
     --output projs \
     --plot-mode PYQTGRAPH
 ````
-(Time on my machine: ~2 minutes, 55 seconds) 
-The option `--plot-mode PYQTGRAPH` will display a plot window of the projection data after it is created. (This will not be available if the module was not installed with the `[gui]` option.)
+(Time on my machine: ~2 minutes, 56 seconds) 
+The options `--ssu 3 --ssv 3` indicate that the line attenuation for each detector pixel is to be calculated as the average of a 3-by-3 grid of exact line integrals. This kind of improved accuracy is necessary for some phantoms, including this one, if using a large detector pixel size. The option `--plot-mode PYQTGRAPH` will display a plot window of the projection data after it is created. (This will not be available if the module was not installed with the `[gui]` option.)
 
 Now, let's reconstruct with three different algorithms:
  1. Global Backprojection-Convolution (GBC)
@@ -148,10 +151,14 @@ aprict-cli recon gbc-cone-cyl \
     --voxwid 0.5 \
     --volshape '(525,525,525)' \
     --alg GREWAR \
+    --f1 1.2 \
+    --f2 6 \
+    --n 9 \
+    --eps 1e-9 \
     --output recon_gbc \
     --plot-mode PYQTGRAPH
 ```
-(Time on my machine: ~37 seconds)
+(Time on my machine: ~39 seconds)
 
 To perform an FBP reconstruction, we do:
 ```bash
@@ -164,7 +171,7 @@ aprict-cli recon fbp-cone-cyl \
     --output recon_fbp \
     --plot-mode PYQTGRAPH
 ```
-(Time on my machine: ~12 seconds)
+(Time on my machine: ~19 seconds)
 
 To perform an FBP-VM reconstruction, we change `--alg` to `NOOGREWAR`:
 ```bash
@@ -177,7 +184,7 @@ aprict-cli recon fbp-cone-cyl \
     --output recon_fbp-vw \
     --plot-mode PYQTGRAPH
 ```
-(Time on my machine: ~22 seconds)
+(Time on my machine: ~26 seconds)
 
 
 </details>
@@ -215,7 +222,7 @@ aprict-cli project cbfp \
     --output projs \
     --plot-mode PYQTGRAPH
 ````
-(Time on my machine: ~1 minute, 46 seconds) 
+(Time on my machine: ~1 minute, 48 seconds) 
 The option `--plot-mode PYQTGRAPH` will display a plot window of the projection data (a sinogram) after it is created. (This will not be available if the module was not installed with the `[gui]` option.) 
 
 You will find a copy of the `ROCK_CAN` phantom in the output directory `projs`, beside the projection data. The phantom is always copied to the output directory, to ensure a good record of the inputs to the simulation.

{aprict-0.0.23 → aprict-0.0.24.dev1}/src/aprict/cb/flatpanel.py

@@ -33,6 +33,7 @@ import numpy.typing as npt  # ruff: ignore[unused-import]
 from .. import logtools
 from ..cuda import header_injection, structs
 from ..cuda.kern_structs import _np_dtype_to_c_type
+from ..lib import gpu_tools
 from ..volume import FBVolume, Volume
 
 _delay_compile: bool = (
@@ -801,6 +802,11 @@ class ProjectionSet:
             )
         return ProjectionSet(self._proj_geom, data=cp.zeros_like(self._data))
 
+    def __del__(self) -> None:  # ruff: ignore[undocumented-magic-method]
+        # HOTFIX to work around a pinned-memory leak in CuPy
+        # Unfortunately, this reduces GPU utilisation.
+        gpu_tools.clear_pinned_cache()
+
 
 class ConfigProject:
     """Configuration object for projections. Some options will have no effect if a Volume or FBVolume."""

{aprict-0.0.23 → aprict-0.0.24.dev1}/src/aprict/lib/ne_tools.py

@@ -162,12 +162,15 @@ def downbin(  # ruff: ignore[complex-structure]
     return xnew
 
 
-def save(path: pathlib.Path, arr: np.ndarray) -> None:
+def save(path: pathlib.Path, arr: np.ndarray, flush: bool = True) -> None:
     """Save an array to a path in a faster manner than np.save."""
     mm: np.memmap = np.lib.format.open_memmap(
         path, shape=arr.shape, dtype=arr.dtype, mode="w+"
     )
     ne.evaluate("x", out=mm, local_dict={"x": arr})
+    if flush:
+        # Write changes to disk now.
+        mm.flush()
 
 
 class NETHREADS(enum.Enum):

{aprict-0.0.23 → aprict-0.0.24.dev1}/src/aprict/logtools/pbar.py

@@ -140,6 +140,11 @@ class ProgressBar:
                 speed_estimate_period=86400,  # period in seconds
             )
 
+    @property
+    def maxval(self) -> int | float:
+        """Maximum value of the progress bar."""
+        return self._maxval
+
     def __enter__(self) -> ProgressBar:  # ruff: ignore[undocumented-magic-method]
         if self._enable:
             self._cm.__enter__()

{aprict-0.0.23 → aprict-0.0.24.dev1}/src/aprict/scripts/recon_cbfp_cyl_fbp.py

@@ -872,7 +872,7 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
             if logger is not None:
                 msg = f"Saving full volume to '{savepath}'..."
                 logger.debug(msg)
-            lib.ne_tools.save(savepath, x)
+            lib.ne_tools.save(savepath, x, flush=True)
             if logger is not None:
                 msg = "Full volume now saved to disk."
                 logger.debug(msg)
@@ -1139,16 +1139,17 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
                 logger.debug(msg)
                 if vol is not None:
                     # Not efficient. Can't copy from CUDAArray directly to the chunk because it's not C-contiguous.
-                    ans = vol.as_numpy()
-                    lib.ne_tools.fastcopy(src=ans, dst=vol_np[..., sl_vol])
-                    del ans  # allow the memory to be freed
-                    lib.gpu_tools.clear_pinned_cache()
+                    vol.as_numpy(
+                        out=vol_np[..., sl_vol], allow_indirect_copy=True
+                    )
+                    del vol
+                    lib.gpu_tools.clear_cache()
             else:
                 msg = "No need to transfer volume chunk from GPU. No work performed."
                 logger.debug(msg)
+                del vol
 
             # Free some memory
-            del vol
             lib.gpu_tools.clear_cache()
             lib.gpu_tools.clear_pinned_cache()
 

{aprict-0.0.23 → aprict-0.0.24.dev1}/src/aprict/scripts/recon_cbfp_cyl_gbc.py

@@ -294,6 +294,12 @@ _progdesc = "Perform an FBP reconstruction from cylindrical flat-panel cone-beam
         default=8080,
         help="Port to use for serving things via a web interface, if applicable.",
     ),
+    ce.option(
+        "--memmap-vols",
+        type=lib.click_tools.BoolArgType(),
+        default=True,
+        help="Memory-map volumes, rather than keeping them in RAM.",
+    ),
 )
 @ce.option_group(
     "resource options",
@@ -317,6 +323,10 @@ def validate_args(args: types.SimpleNamespace) -> None:
         msg = f"Output path '{args.output!s}' already exists."
         raise ce.UsageError(msg)
 
+    if args.output is None and args.memmap_vols:
+        msg = "Cannot use --memmap-vols True while --output is not specified."
+        raise ce.UsageError(msg)
+
     # If outputting intermediates, then the user must also specify the --output directory
     if args.output_intermediates and not args.output:
         msg = "Cannot output intermediate tomograms when no output directory was specified. Specify an output directory with --output path/to/new/dir"
@@ -903,9 +913,20 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
 
     msg = "Allocating and zeroing the backprojection arrays..."
     logger.info(msg)
-    # Put the z axis first, so that we can slice it in Z and have a C-contiguous array. This is necessary to copy directly from a cupy array to a numpy array.
-    vol_np = np.empty(padded_volshape, dtype=vol_dtype)
-    volmini_np = np.empty(padded_volshape_mini, dtype=vol_dtype)
+    if args.memmap_vols:
+        assert path_output is not None  # ensured by validate_args(...)
+        # These are temporary arrays which may be overwritten during filtration.
+        savepath = path_output.joinpath("tmp_backproj" + ".npy")
+        vol_np = np.lib.format.open_memmap(
+            savepath, shape=padded_volshape, dtype=vol_dtype, mode="w+"
+        )
+        savepath = path_output.joinpath("tmp_backproj" + "_mini" + ".npy")
+        volmini_np = np.lib.format.open_memmap(
+            savepath, shape=padded_volshape_mini, dtype=vol_dtype, mode="w+"
+        )
+    else:
+        vol_np = np.empty(padded_volshape, dtype=vol_dtype)
+        volmini_np = np.empty(padded_volshape_mini, dtype=vol_dtype)
     lib.ne_tools.set_array_to_constant_pbar(
         vol_np,
         const=0,
@@ -923,11 +944,18 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
         pbar_persist=False,
     )
 
+    # Convert the progress display ta more intuitive set of units using the progratio.
+    # The displayed progress will correspond to bytes in the padded backprojection.
+    progratio: float = (
+        math.prod(padded_volshape)
+        * np.dtype(vol_dtype).itemsize
+        / bytes_cumsum[-1]
+    )
     curr_slice: slice | None = None
     logger.info("Beginning reconstruction...")
     with logtools.ProgressBar(
         desc="Backproj.:",
-        maxval=int(bytes_cumsum[-1]),
+        maxval=math.ceil(bytes_cumsum[-1] * progratio),
         logger=logger,
         persist=True,
         unit_is_bytes=True,
@@ -1042,7 +1070,8 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
                         prog_evs_bytes.append(
                             (
                                 ev,
-                                (
+                                progratio
+                                * (
                                     (
                                         (sl.stop - sl.start)
                                         / tot_num_nonskipped_projs
@@ -1075,8 +1104,6 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
             if tot_num_nonskipped_projs == 0:
                 msg = "No need to transfer volume chunk from GPU. No work performed."
                 logger.debug(msg)
-                del vol_weightsum
-                del minivol_weightsum
             else:
                 # Normalise to the correct weighting.
                 # This will multiply both a numerator and a divisor by the same amount.
@@ -1120,33 +1147,36 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
                     minivol.divide_softzero(minivol_weightsum, eps=args.eps)
                     del vol_weightsum
                     del minivol_weightsum
+                    lib.gpu_tools.clear_cache()
+                    lib.gpu_tools.clear_pinned_cache()
 
                 # Save the backprojections into the arrays
                 msg = "Transferring backprojected volume chunk from GPU..."
                 logger.debug(msg)
+                # Copy minivol.
+                # Not efficient. Can't copy from CUDAArray directly to the chunk because it's not C-contiguous.
+                minivol.as_numpy(
+                    out=volmini_np[..., sl_minivol], allow_indirect_copy=True
+                )
+                del minivol
+                lib.gpu_tools.clear_cache()
+                # Copy vol.
                 if vol is not None:
                     # Not efficient. Can't copy from CUDAArray directly to the chunk because it's not C-contiguous.
-                    ans = vol.as_numpy()
-                    lib.ne_tools.fastcopy(src=ans, dst=vol_np[..., sl_vol])
-                    del ans  # allow the memory to be freed
-                    lib.gpu_tools.clear_pinned_cache()
-                # Not efficient. Can't copy from CUDAArray directly to the chunk because it's not C-contiguous.
-                ans = minivol.as_numpy()
-                lib.ne_tools.fastcopy(src=ans, dst=volmini_np[..., sl_minivol])
-                del ans  # allow the memory to be freed
-                lib.gpu_tools.clear_pinned_cache()
-
-            # Free some memory
-            del vol, minivol
-            lib.gpu_tools.clear_pinned_cache()
-            lib.gpu_tools.clear_cache()
+                    vol.as_numpy(
+                        out=vol_np[..., sl_vol], allow_indirect_copy=True
+                    )
+                    del vol
+                    lib.gpu_tools.clear_cache()
 
-        pbar.update_to(int(bytes_cumsum[-1]))
+        pbar.update_to(pbar.maxval)
 
     msg = "Filtering volumes..."
     logger.info(msg)
     with rich_console.status(msg, spinner="dots"):
         # Filter the mini recon first.
+        msg = "Filtering padded minivol..."
+        logger.debug(msg)
         minivol_voxwid = args.voxwid * args.n
         ndarray_container = lib.np_tools.NDArrayContainer(array=volmini_np)
         minivol_padded_filtered_np = gbc.filters.ramp_filter_conebeam(
@@ -1167,16 +1197,20 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
             ]
         )
         del volmini_np  # not going to reuse this, so delete this reference so the filtration can free some memory.
+        msg = "Filtering unpadded minivol..."
+        logger.debug(msg)
         minivol_unpadded_filtered_np = gbc.filters.ramp_filter_conebeam(
             ndarray_container,
             minivol_voxwid,
             2 * vartheta_v,
             theta_soft=args.softspan_v,
             auto_dc_adjust=False,
-            allow_overwrite_arr=True,  # we're not going to reuse the 'minivol_cp' object after this.
+            allow_overwrite_arr=True,  # we're not going to reuse the 'volmini_np' object after this.
             logger=logger,
             copy_to_array=None,
         )
+        msg = "Producing minivol correction array..."
+        logger.debug(msg)
         minivol_correction_np = (
             minivol_padded_filtered_np[
                 mvc_padx : (-mvc_padx) or None,
@@ -1192,6 +1226,8 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
             ndarray_container,
         )
         # Filter the full recon now
+        msg = "Filtering padded vol and adding upscaled minivol correction..."
+        logger.debug(msg)
         ndarray_container = lib.np_tools.NDArrayContainer(array=vol_np)
         vol_voxwid = args.voxwid
         del vol_np  # allows memory to be freed, because the gbc.filters.ramp_filter_conebeam is free to overwrite the ndarray_container.array
@@ -1201,7 +1237,7 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
             2 * vartheta_v,
             theta_soft=args.softspan_v,
             auto_dc_adjust=False,
-            allow_overwrite_arr=True,
+            allow_overwrite_arr=True,  # not going to need the 'vol_np' object after this.
             logger=logger,
             copy_to_array=None,  # Should be a contiguous array. E.g., not an arbitrary view.
             lowres_correction_array=minivol_correction_np,
@@ -1231,6 +1267,8 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
             f"Estimated DC component to be subtracted: {dc_component:.5g}"
         )
         # Correct the DC component
+        msg = "Adding DC correction..."
+        logger.debug(msg)
         ne.evaluate(
             "v - dcc",
             out=vol_filtered_np,
@@ -1250,6 +1288,7 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
 
         # Clean up
         lib.gpu_tools.clear_cache()
+        lib.gpu_tools.clear_pinned_cache()
 
     if path_output is not None:
         msg = "Saving volume to disk..."
@@ -1258,9 +1297,17 @@ def main_from_args(args: types.SimpleNamespace) -> None:  # ruff: ignore[complex
             x = vol_filtered
             # Save the volume as a .npy file
             tomo_path = path_output.joinpath("recon.npy")
-            lib.ne_tools.save(tomo_path, x)
+            lib.ne_tools.save(tomo_path, x, flush=True)
             msg = "Full volume now saved to disk."
             logger.debug(msg)
+            # Now get rid of the temporary files, if applicable
+            if args.memmap_vols:
+                savepath = path_output.joinpath("tmp_backproj" + ".npy")
+                savepath.unlink()  # delete file
+                savepath = path_output.joinpath(
+                    "tmp_backproj" + "_mini" + ".npy"
+                )
+                savepath.unlink()  # delete file
             # Save slices
             for axis, axis_str in enumerate("xyz"):
                 msg = f"Saving '{axis_str}' slice..."

{aprict-0.0.23 → aprict-0.0.24.dev1}/src/aprict/scripts/recon_cbfp_landweber.py

@@ -521,7 +521,7 @@ def main() -> None:
             save_path.joinpath("tomograms") if is_intermediate else save_path
         )
         tomo_path.mkdir(exist_ok=True)
-        lib.ne_tools.save(tomo_path.joinpath(id_str + ".npy"), x)
+        lib.ne_tools.save(tomo_path.joinpath(id_str + ".npy"), x, flush=True)
 
         # Save midslices from volume as .png files
         # The method used here can be a bit slow, because I get a full numpy array and process that.

{aprict-0.0.23 → aprict-0.0.24.dev1}/src/aprict/version.py

@@ -16,4 +16,4 @@
 # This file is read from by 'setup.py', from '__init__.py', and also by a git hook which increments the dev number each time a commit is made on the 'dev' branch. There should only be one line in the file which begins with "version".
 """File which stores the current version number of this module."""
 
-version = "0.0.23"
+version = "0.0.24.dev1"

{aprict-0.0.23 → aprict-0.0.24.dev1}/src/aprict/volume.py

@@ -193,8 +193,21 @@ class Volume[
             self._texture, vol._texture, tout=self._texture, eps=eps
         )
 
-    def as_numpy(self, out: np.ndarray | None = None) -> np.ndarray:
+    def as_numpy(
+        self, out: np.ndarray | None = None, allow_indirect_copy: bool = False
+    ) -> np.ndarray:
         """Copy this volume's data into a numpy array."""
+        # Note: the option 'allow_indirect_copy' allows a fast-but-not-fastest copy when the output array is not C-contiguous.
+        if out is not None and not out.flags["C_CONTIGUOUS"]:
+            # The output volume is not C contiguous. A copy cannot be performed unless
+            # allow_indirect_copy=True is supplied.
+            if not allow_indirect_copy:
+                msg = "The array supplied to the out=... keyword is not C contiguous. The data cannot be copied directly, and indirect copying must be explicitly allowed via allow_indirect_copy=True."
+                raise ValueError(msg)
+            out_ = self.as_numpy(out=None)
+            ne_tools.fastcopy(src=out_, dst=out)
+            gpu_tools.clear_pinned_cache()
+            return out
         x = (
             out
             if out is not None
@@ -209,7 +222,7 @@ class Volume[
     def save(self, path: pathlib.Path) -> None:
         """Save this volume into a numpy array on disk."""
         x = self.as_numpy()
-        ne_tools.save(path, x)
+        ne_tools.save(path, x, flush=True)
 
     def as_cupy(self, out: cp.ndarray | None = None) -> cp.ndarray:
         """Copy this volume's data into a cupy array."""

{aprict-0.0.23 → aprict-0.0.24.dev1/src/aprict.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aprict
-Version: 0.0.23
+Version: 0.0.24.dev1
 Summary: Tools to simulate and reconstruct XCT transmission data.
 Author: Murdock Grewar
 Author-email: Murdock Grewar <murdock.grewar@anu.edu.au>
@@ -36,6 +36,10 @@ Requires-Dist: neuroglancer>=2.41.0; extra == "gui"
 Dynamic: author
 Dynamic: license-file
 
+[![Static Badge](https://img.shields.io/badge/ORCiD-0000--0002--0015--8284-green?logo=orcid)](https://orcid.org/0000-0002-0015-8284)
+[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.23632625.svg)](https://doi.org/10.5281/zenodo.23632625)
+![PyPI Version](https://img.shields.io/pypi/v/aprict)
+[![Project Status: WIP – Initial development is in progress, but there has not yet been a stable, usable release suitable for the public.](https://www.repostatus.org/badges/latest/wip.svg)](https://www.repostatus.org/#wip)
 <details open><summary><h1>About</h1></summary>
 
 Description
@@ -71,8 +75,7 @@ Known limitations
   * Volume size is limited by CUDA texture limit. For example, a `(2400,2400,4800)` volume can be reconstructed, but not a `(30000,30000,1)` volume, due to a volume of this size not fitting within a single CUDA texture
   * The FORBILD shape `Tetrahedron` is not yet implemented, and no primitive shape is permitted to have more than 3 clipping planes applied, due to internal limitations. All other FORBILD phantoms should be supported.
   * Only one GPU is utilised by this module, at the present time. There is no defined behaviour for which GPU will be chosen.
-  * Not fully optimised: among other issues, 
-    * the GBC reconstruction script assumes that the whole volume (or more) can fit in RAM. This limitation may be lifted in the future. (Note: a present workaround is to create a large swapfile on disk, since swap space is treated as RAM when the physical RAM is exhausted.)
+  * Not fully optimised
   * CPU and GPU work not completely asynchronous: due to unresolved bugs, various asynchronous GPU operations have been disabled. This may not have much impact on large reconstructions, but could result in poor GPU utilisation for smaller reconstructions.
   * No smart batching of (back)projections: smarter code would analyse the geometry of projections, and batch together (back)projections that involve similar regions of the volume. Currently, the code does not do any kind of clever batching like this, and so the (back)projection code is not as performant as it could be.
 
@@ -99,7 +102,7 @@ You can install `aprict` like you would any python package, using `pip install`.
 Example installation using `uv`: 
 ```bash
 $ uv venv -p 3.14
-$ uv pip install aprict[gui]==0.0.19
+$ uv pip install aprict[gui]==0.0.24
 $ source .venv/bin/activate
 ```
 You should now have several executables available on `$PATH`. Try `aprict-cli --help` to see the polished commandline interface. 
@@ -118,11 +121,11 @@ Activate the environment:
 ```bash
 source .venv/bin/activate
 ```
-Now, try the shell command `aprict-cli --help`. Here is the output as at `v0.0.19`:
+Now, try the shell command `aprict-cli --help`. Here is the output as at `v0.0.24`:
 ```
 Usage: aprict-cli [OPTIONS] COMMAND [ARGS]...
 
-  Entry point for aprict CLI.
+  Entry point for CLI interface to aprict version 0.0.24
 
 Options:
   --version   Show the version and exit.
@@ -170,8 +173,8 @@ aprict-cli project cbfp \
     --output projs \
     --plot-mode PYQTGRAPH
 ````
-(Time on my machine: ~2 minutes, 55 seconds) 
-The option `--plot-mode PYQTGRAPH` will display a plot window of the projection data after it is created. (This will not be available if the module was not installed with the `[gui]` option.)
+(Time on my machine: ~2 minutes, 56 seconds) 
+The options `--ssu 3 --ssv 3` indicate that the line attenuation for each detector pixel is to be calculated as the average of a 3-by-3 grid of exact line integrals. This kind of improved accuracy is necessary for some phantoms, including this one, if using a large detector pixel size. The option `--plot-mode PYQTGRAPH` will display a plot window of the projection data after it is created. (This will not be available if the module was not installed with the `[gui]` option.)
 
 Now, let's reconstruct with three different algorithms:
  1. Global Backprojection-Convolution (GBC)
@@ -186,10 +189,14 @@ aprict-cli recon gbc-cone-cyl \
     --voxwid 0.5 \
     --volshape '(525,525,525)' \
     --alg GREWAR \
+    --f1 1.2 \
+    --f2 6 \
+    --n 9 \
+    --eps 1e-9 \
     --output recon_gbc \
     --plot-mode PYQTGRAPH
 ```
-(Time on my machine: ~37 seconds)
+(Time on my machine: ~39 seconds)
 
 To perform an FBP reconstruction, we do:
 ```bash
@@ -202,7 +209,7 @@ aprict-cli recon fbp-cone-cyl \
     --output recon_fbp \
     --plot-mode PYQTGRAPH
 ```
-(Time on my machine: ~12 seconds)
+(Time on my machine: ~19 seconds)
 
 To perform an FBP-VM reconstruction, we change `--alg` to `NOOGREWAR`:
 ```bash
@@ -215,7 +222,7 @@ aprict-cli recon fbp-cone-cyl \
     --output recon_fbp-vw \
     --plot-mode PYQTGRAPH
 ```
-(Time on my machine: ~22 seconds)
+(Time on my machine: ~26 seconds)
 
 
 </details>
@@ -253,7 +260,7 @@ aprict-cli project cbfp \
     --output projs \
     --plot-mode PYQTGRAPH
 ````
-(Time on my machine: ~1 minute, 46 seconds) 
+(Time on my machine: ~1 minute, 48 seconds) 
 The option `--plot-mode PYQTGRAPH` will display a plot window of the projection data (a sinogram) after it is created. (This will not be available if the module was not installed with the `[gui]` option.) 
 
 You will find a copy of the `ROCK_CAN` phantom in the output directory `projs`, beside the projection data. The phantom is always copied to the output directory, to ensure a good record of the inputs to the simulation.